Post step: stop writing the multi-pack-index during mirror maintenance

[piob-io]

Summary

Since v1.8.0 the post step runs git repack -d -l -n --geometric=2 --write-midx on every job. On large mirrors (a few GiB, millions of objects) the fleet shows a bimodal post step: most runs finish maintenance in well under a second, but a recurring ~25% take 6–16 s inside that repack command with the same inputs (same two kept packs, nothing deferred, one small incoming pack). The roll-up itself is tiny in those runs; what scales with the mirror is --write-midx: git rewrites the multi-pack-index whole each time, reading every pack's .idx (tens of MiB per kept pack) and writing about as many bytes again, and on a freshly mounted sticky disk those reads are cold.

Local measurement on a 2.8M-object / 1.3 GiB two-kept-pack fixture with one small incoming pack, page cache dropped before each run: --geometric=2 --write-midx 0.55–0.59 s, --geometric=2 alone ~0.01 s; same ratio on a 1M-object fixture. The end-to-end confirmation on a real staging sticky disk is the steady-state legs added to the checkout-mirror-maintenance-e2e workflow in test-workflows (baseline main vs this branch on the same seeded mirror).

Change (runMirrorMaintenance):

-  repackArgs = ['-a', '-d', '-l', '-n', '--write-midx']        // reclaim
+  repackArgs = ['-a', '-d', '-l', '-n']
-  repackArgs = ['-d', '-l', '-n', '--geometric=2', '--write-midx']  // incremental
+  repackArgs = ['-d', '-l', '-n', '--geometric=2']
+  if (!(await removeMultiPackIndex(mirrorPath))) return fail(false, ...)  // before the repack

Lookups across the handful of packs the geometric repack leaves (kept base packs + one or two roll-ups) are fast without a midx, and bitmaps were already off. The index a v1.8.0 run left on existing mirrors is removed up front rather than left to go stale; git repack -d would drop it itself on the first run that folds something, but not on a no-op run.

Everything else is unchanged: .keep selection and byte bound, the 14-day reclaim, the fail-open commit, stale-file cleanup.

Tests: existing real-git tests now assert no midx is present after incremental and reclaim runs; new test seeds a midx (as an earlier version would have) and checks it is removed and every pack it listed is still rolled up, fsck clean. cleanup test asserts --write-midx is not passed.

Link to Devin session: https://app.devin.ai/sessions/a3b42ef15226438faf3f9cdde9662918
Open in Devin Desktop: https://app.devin.ai/desktop/session/a3b42ef15226438faf3f9cdde9662918?variant=devin
Requested by: @piob-io

---

View with [code]smith Autofix with [code]smith
^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled. (Staging)}

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring